US5418915A - Arithmetic unit for SIMD type parallel computer - Google Patents

Arithmetic unit for SIMD type parallel computer Download PDF

Info

Publication number
US5418915A
US5418915A US08251651 US25165194A US5418915A US 5418915 A US5418915 A US 5418915A US 08251651 US08251651 US 08251651 US 25165194 A US25165194 A US 25165194A US 5418915 A US5418915 A US 5418915A
Authority
US
Grant status
Grant
Patent type
Prior art keywords
instruction
arithmetic
address
parallel computer
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
US08251651
Inventor
Motohiko Matuda
Taiichi Yuasa
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sumitomo Metal Industries Ltd
Original Assignee
Sumitomo Metal Industries Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Grant date

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30181Instruction operation extension or modification
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F15/00Digital computers in general; Data processing equipment in general
    • G06F15/76Architectures of general purpose stored program computers
    • G06F15/80Architectures of general purpose stored program computers comprising an array of processing units with common control, e.g. single instruction multiple data processors
    • G06F15/8007Architectures of general purpose stored program computers comprising an array of processing units with common control, e.g. single instruction multiple data processors single instruction multiple data [SIMD] multiprocessors
    • G06F15/8023Two dimensional arrays, e.g. mesh, torus
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3885Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3885Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units
    • G06F9/3887Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units controlled by a single instruction, e.g. SIMD

Abstract

In the SIMD parallel computer according to the present invention, the arithmetic unit has an instruction storage means for storing a local instruction, a selecting means for selecting any one of the common instruction and the local instruction stored in the instruction storage means, a selection control means for controlling a selecting operation of the selecting means and a means for executing the selected instruction. In an arrangement of the SIMD parallel computer of this invention is that each arithmetic part is connected to the memory via an address changeover circuit and a data changeover circuit as well. One arithmetic part is selectively connected to the memory under the selection control handled by the central control circuit. Further in the parallel computer according to the present invention for performing parallel processing of the data through communications between the plurality of processors arrayed in grids, each processor has output links extending two directions of a first diagonal line of the grid and also input links extending in two directions of a second diagonal line thereof. The output links are connected to input or output links of 3-neighbor processors. Provided is a communication circuit for selecting one of the bidirectional output links by 1 bit of a 2-bit direction signal given to establish the communication direction and one of the bidirectional input links by another 1-bit.

Description

This application is a continuation of application Ser. No. 07/741,568, filed Aug. 7, 1991, now abandoned.

BACKGROUND OF THE INVENTION

The present invention relates to a parallel computer having a Single Instruction Stream/Multiple Data Stream (SIMD) architecture for executing a single instruction with respect to multiple data by use of a plurality of arithmetic units, and more particularly, to an architecture of processing elements (PEs) of the SIMD computer as well as to a communication connection network between the PEs.

A Multiple Instruction Stream/Multiple Data Stream (MIMD) architecture and a Single Instruction Stream/Multiple Data Stream (SIMD) architecture are taxonomically exemplified as typical architectures of parallel computers, especially of hyperparallel computers for executing parallel operations by arranging several hundreds to several ten thousands of arithmetic units in parallel.

The MIMD parallel computer is conceived as a parallel computer of a system such that a plurality of arithmetic units are each controlled by a string of particular instructions. Each arithmetic unit has a high independency, and hence a generality in purpose of the system is enhanced. Complicated parallel processing can therefore be performed. This type of parallel computer tends, however, to cause an increase in overhead for communications and synchronizations between the strings of instructions which are working in the respective arithmetic units; and the control is also liable to be complicated.

The SIMD parallel computer is conceived as a parallel computer of a system such that the plurality of arithmetic units are controlled by a single string of instructions. All the arithmetic units are synchronized and operated by the single string of instructions. Hence, control is simple, and there is no necessity for giving instructions to respective arithmetic units. Therefore, this type of parallel computer may be relatively easily scaled up and suited to such simple numeric processing and image processing that simple arithmetic operations are repeatedly applied to a large amount of data.

Because of the characteristics described above, the SIMD parallel computers are dominant as commercially available parallel computers which are now utilized. Connection Machine (CM2) of Thinking Machine Corp. and MP-1 of Mas Par Corp. are examples of the SIMD parallel computers.

CM2 is a hyperparallel machine based on such configuration that 16 pieces of 1-bit processors are integrated on one LSI chip, and 65536 processors are hypercube-connected on a 4096-chip unit. Each chip incorporates routers in addition to the processors. A memory chip is externally attached, and a memory capacity per processor is 256K bits. Inter-processor communication connections are based on grid connection (NEWS grid) communication. A communication mode known as a direct hypercube communication is available other than the typical message address based communication. The hypercube communication is effected by directly connecting 12 processors among 16 processors on the chip through 12 lengths of communication links.

MP-1 is a hyperparallel computer which has attained a parallelism as high as 16284 (128×128) on the basis of a multi-stage, i.e., 3-stage connection network using 64×64 cross bar switch LSIs and two-dimensional grid connections (X-Net) for 8 pieces of adjacent processors by use of 512 CMOS LSIs each obtained by integrating 32 pieces of 4-bit processors on one chip. MP-1 has a 16 k-bit local memory per processor.

FIG. 8 is a block diagram showing an example of architecture of a conventional SIMD parallel computer.

Designated at 101 in FIG. 8 is a control circuit for handling the whole control over the parallel computer. The control circuit 101 generates a control signal CTR, an instruction INS and an address ADD. The control circuit 101 imparts the generated control signal CTR, the same instruction INS and the same address ADD to a plurality of arithmetic units 102, 102 . . . arranged in parallel via a control line 111, an instruction bus 112 and an address bus 113.

In the SIMD parallel computer, all the processors typically execute the same string of instructions issued from the control unit. However, when requiring ununiform calculations as seen under boundary conditions associated with, e.g., a particle motion and a thermal flow, arithmetic contents of some processors differ from those of a majority of other processors in some cases. In some of the prior art parallel computers, on this occasion the contents presumable in all cases have hitherto been given to the strings of instructions, the instructions are executed by operation authorizing flags provided in the respective PEs; or alternatively whether the execution is skipped or not is determined. In, for instance, CM2 each processor normally executes the same arithmetic operation. However, the execution can be skipped depending on an inside status thereof.

FIG. 9 is a block diagram showing a construction of the arithmetic units 102 of the conventional SIMD parallel computer. The instruction INS sent from the control circuit 101 via the instruction bus 112 is given to the arithmetic element 122.

The arithmetic element 122 is constructed to selectively execute a plurality of arithmetic processes such as addition and subtraction. The arithmetic element 122 performs an arithmetic process corresponding to the given instruction.

The address ADD supplied from the control circuit 101 via the address bus 113 is provided to the local memory 123 or the register group 121. The local memory 123 stores arithmetic data used in the arithmetic element 122 and data on the arithmetic result. The register group 121 temporarily stores the arithmetic data stored in the local memory 123 and the data in the middle of arithmetic operation and supplies the data to the arithmetic element 122. Besides, the register group 121 has an area for storing an operation authorizing flag 124 which authorizes the arithmetic unit 102 in terms of operation depending on a status thereof. A status of this operation authorizing flag 124 is controlled by the control signal CTR transmitted via the control line 111 of the control circuit 101.

In the thus constructed conventional SIMD parallel computer, the arithmetic operation based on the same instruction INS is effected referring to the same address ADD of the local memory 123 at the same point of time in all the arithmetic units 102, 102 . . . .

The operation of each arithmetic unit 102 is controllable depending on the status of the operation authorizing flag 124. More specifically,the operation authorizing flag is brought into an unauthorized status, whereby an execution of a certain instruction can be skipped per arithmetic unit 102. A flexibility of calculation is thereby obtainable. This also makes it possible to cause the SIMD parallel computer to work as if a MIMD parallel computer does.

To perform the arithmetic processes pursuant to different instructions per arithmetic unit in the conventional single instruction parallel computer, however, it is required that the following operations be effected the same number of times as the number of instructions thereof. Namely, the operation authorizing flag is put into an authorized status for only the arithmetic unit which is to execute a certain instruction. Only the authorized arithmetic unit executes a desired arithmetic process, while other arithmetic units skip that instruction and perform no arithmetic process.

With this arrangement, it is feasible to execute a sequence of instructions different from the whole in some PEs. Skipping of instruction, however, produces a dead time in terms of processing and is therefore unfavorable. Hence, what is desired is not skipping of instructions but an architecture that permits the executions of the different instructions.

Besides, the SIMD parallel computer generally needs a memory having a capacity large enough to store the data to be processed into each arithmetic unit. For this reason, the memory is ordinarily, though the arithmetic units are actualized on one LSI chip, provided outwardly of the LSI. When actualizing the arithmetic unit on one LSI chip, a plurality of arithmetic units (CPU) are formed as one arithmetic module on one LSI.

FIG. 10 is a block diagram showing an example of architecture of the above-mentioned conventional SIMD parallel computer.

Referring to FIG. 10, the reference numeral 201 represents an arithmetic unit composed of arithmetic parts 211 and memory parts 212. This computer is defined as a parallel computer, so that a multiplicity of arithmetic units 201 are disposed in parallel. A plurality of arithmetic parts 211 that are components of the arithmetic unit 201 are formed on one LSI chip 202. A plurality of such LSI chips 202 are further connected.

Note that each memory part 212 is not formed on the LSI chip 202 but is externally attached. The reason for this is that a large memory capacity is required to be prepared in the SIMD parallel computer, and hence using a dedicated memory circuit for the memory part 212 is more advantageous in terms of many aspects.

The SIMD parallel computer includes a central control circuit 203 for supplying a common memory address and instruction to all the arithmetic units 201. The instruction is issued from this central control circuit 203 to the arithmetic parts 211 of each arithmetic unit 201. The address is also given to the memory parts 212. Each arithmetic part 211 reads the data from the corresponding memory part 212 and executes the arithmetic operation. The result thereof is written as data to the memory part 212.

For the purpose of eliminating such a constraint that the memory address is common to all the arithmetic units among constraints with which the SIMD parallel computer is burdened, an address generating circuit is individually, provided in each arithmetic unit 201.

FIG. 11 is a block diagram showing one example of the construction described above. To be specific, each arithmetic unit 201 has an address generating/changeover circuit 2132 so newly provided on the LSI chip 202 as to be attached to the arithmetic part 211. This address generating/changeover circuit 213 accesses the memory part 212 by, e.g., register indirect addressing on the basis of a memory address given from the central control circuit 203.

Incidentally, in the SIMD parallel computer, a memory data width is, as in the case of CM2 for instance, approximately 1 bit per arithmetic unit. This aims at regulating an increment in the number of outside pins of the LSI chip 202 when forming the multiplicity of arithmetic parts 211 on one LSI chip 202. It is, however, required that an address width for the memory access be adaptive to a capacity of the memory part 212. Specifically, an address width needed is normally 10 bits through 20 bits or larger. Under such circumstances, the arrangement that the multiplicity of arithmetic parts 211 are formed on the LSI chip 202 is restricted in terms of the number of pins of the LSI chip 202.

The following is an explanation of another conceivable method. As illustrated in FIG. 12, the memory address is converted from a parallel signal into a series signal by means of a P/S (parallel/series) converting circuit 214 for converting the memory address into the series signal. The series signal is outputted outwardly of the LSI chip 202. The signal is restored to the original parallel signal by an S/P (series/parallel) converting circuit 215. The parallel signal is then inputted to the memory 212. There arises, however, a problem inherent in this method, wherein it needs an extra time for the conversion of the memory address into the series signal and the restoration to the parallel signal and further a circuit therefor, resulting in an increase in costs.

As discussed above, the arithmetic unit is divided into the arithmetic parts and the memory parts, and the plurality of arithmetic parts alone are formed en bloc on one chip (LSI) in the SIMD parallel computer. Based on this construction, when effecting register indirect addressing while referring to the memory part at an address indicated by a register within the arithmetic part, it is necessary for the arithmetic units to individually output the addresses. In this case, the address needs a bit width corresponding to the memory capacity of the memory part. When the plurality of arithmetic parts are formed en block on one chip, however, it is impossible to secure the necessary address width described above in terms of a restriction in the number of pins of the chip.

Concretely, in the great majority of LSI-based single instruction parallel computers, the address width is approximately 10 bits through 20 bits. Formed are 16 through 128 pieces of arithmetic parts per chip. Hence, the necessary number of pins for the memory addresses is 160 at the minimum. Those pins can not be packaged.

Furthermore, the SIMD parallel computer adopts configurations such as a grid connection communication links (NEWS grid) which typically provide connections in grids for communications between adjacent PEs and two-dimensional grid connections (X grid).

FIG. 13 is a block diagram illustrating an NEWS grid connection network of the conventional parallel computer. Processors 511, 511 . . . arrayed in grids are connected to east, west, south and north (E, W, S, N) grid-4-neighbor processors 511, 511 . . . via bidirectional communication links.

The following are demands for the communication connections to the grid-4-neighbor processors in the parallel computer where the processors are arrayed in grids.

(1) The data is transmitted in a selected direction among the four directions E, W, S and N of the grid. The target processor to which the data is transmitted receives data from a direction opposite thereto. For example, when a certain processor transmits the data in the direction N. The N-directional target processor receives the data from the direction S. The communication is thus established.

(2) The communication can be effected in the selected direction among the four directions of the grid.

(3) The number of communication links is small.

(4) All the processors simultaneously perform the communications in the same direction.

To meet those demands, the conventional parallel computer takes such an arrangement that the processors in the directions E, W, S and N of the grid are directly connected through bidirectional links.

FIG. 14 is a block diagram showing a construction of a conventional communication circuit of each of the processors 511, 511 . . . . The processors 511, 511 . . . include arithmetic elements 512 and the communication circuits 513. The arithmetic element 512 effects a process on the data received via the communication circuit 513, the process being pursuant to an instruction outputted from a control circuit (not shown) for handling the whole control. The processed data is transmitted through the communication circuit 513. The control circuit supplies 2-bit direction signal DS ti the communication circuit 513 of the processors 511, 511 . . . .

The communication circuit 513 consists of a 2-to-4 decoder 17 for decoding the 2-bit direction signal DS into four signals indicating the four directions N, E, W and S, output buffers 514a-514d or input buffers 515a-515d for respectively selecting the directions the transmission or receive data in response to the four decoded signals and an OR gate 516 for giving the receive data to the arithmetic element 512.

The 2-to-4 decoder 517 decodes the 2-bit direction signal DS into, e.g., 4-direction signals shown in Table 1.

              TABLE 1______________________________________Direction signal           Direction______________________________________00              N → S01              E → W10              W → E11              S → N______________________________________

The output buffers 514a-514d are constructed by using open collector type NAND gates. The four signals transmitted from the 2-to-4 decoder 17 are supplied respectively to ends of input terminals of the output buffers. The transmission data are provided to the other ends thereof.

The transmission data outputted from the output buffers 514a-514d are outputted in any one direction selected among the four directions such as N→S, E→W, W→E and S→N via the bidirectional communication links.

The receive data inputted via the communication links are Inversion-inputted to one ends of the input buffers 515a-515d. The input buffers 515a-515d are constructed by use of AND gates. The four signals are supplied respectively to the other ends of the input buffers. The input buffers 515a-515d selectively output the receive data in response to the four signals; and outputs thereof are given to the arithmetic element 512 through the OR gate 516.

In the prior art parallel computer having the above-described architecture, when the communications are carried out at one time in the direction, e.g., N, the control circuit supplies all the processors 511, 511 . . . with the direction signals DS=00. Only the N→S direction signal of the 2-to-4 decoder thereby becomes 1. Only the output buffer 514a to the direction N and the input buffer 515a from the direction S become conductive, whereby a communicable state is developed. From this state, the transmission data is transmitted in the direction N, while the receive data is received from the direction S.

The conventional parallel computer, however, requires the I/O buffers in every direction. It is also necessary to generate the four signals for specifying the respective directions by decoding the 2-bit direction signal. This presents a problem where the hardware architecture of the communication circuits is intricate.

Two communication links are needed per processor, although each processor uses the communication links in common to the grid-4-neighbor processors.

An X-Net grid structure employed, e.g., MP-1 is obtained by expanding the NEWS grid connection network from the grid-4-neighbor connections to grid-8-neighbor connections. This X-Net grid structure is demonstrated by FIG. 15. As obvious from the Figure, each processor has the bidirectional communication links extending in directions such as north east (NE), north west (NW), south east (SE) and south west (SW) of the grid. NE, NW, SE and SW communication links led from each of four adjacent processors are wired-OR-connected, whereby the adjacent 8-neighbor processors are communicable with each other. With this arrangement, the communications with the 8-neighbor processors can be effected through a relatively small number of communication links. The construction of the communication circuit is, however, still the same as that of FIG. 14 (the I/O links are changed from N, E, W, S to NE, NW, SE, SW). It will be apparent that the communication control becomes more complicated.

SUMMARY OF THE INVENTION

It is a primary object of the present invention, which has been devised under such circumstances explained above, to provide an SIMD parallel computer capable of processing, as in the way with an MIMD parallel computer, a plurality of instructions at a high speed while keeping hardware characteristics of being simple in construction and low in costs by incorporating an instruction register for storing a local instruction into an arithmetic unit and selectively executing the local instruction and a common instruction issued from a control circuit.

It is another object of the present invention to provide an SIMD parallel computer capable of setting address and data bit widths sufficiently without undergoing a constraint of the number of pins of a chip.

It is still another object of the present invention to provide a parallel computer capable of attaining a communication function equivalent to the conventional one between grid-4-neighbor processors with a simple construction by reducing the number of communication links and a quantity of hardware within the communication circuits.

In the SIMD parallel computer according to the present invention, a common instruction is given from a control circuit to a plurality of arithmetic units each including an arithmetic element and a memory, and the given instruction is executed in parallel. In this SIMD parallel computer, the arithmetic unit has an instruction storage means for storing a local instruction, a selecting means for selecting any one of the common instruction and the local instruction stored in the instruction storage means, a selection control means for controlling a selecting operation of the selecting means and a means for executing the selected instruction.

In the computer system of this invention exhibiting such characteristics, when the arithmetic unit executes a plurality of instructions in parallel, the instruction storage means stores the local instruction, and the selecting means is controlled per arithmetic unit so as to select any one of the local instruction and the common instruction. The selected instruction is then executed.

For example, when executing add and subtract instructions in parallel, the add instruction defined as a common instruction is given from the control circuit to each arithmetic unit. At this time, the subtractive instruction is stored in the instruction storage means of the arithmetic unit which performs the subtractive instruction defined as a local instruction. The selecting means of the arithmetic unit for effecting the add instruction is so controlled by the selection control means as to select the add instruction. Conversely, the selecting means of the arithmetic unit for effecting the subtractive instruction is controlled to select the subtractive instruction in the instruction storage means. With this arrangement, the plurality of instructions can be simultaneously executed in parallel in the SIMD parallel computer. It follows that the processing time can be reduced.

In the prior art, the memories are connected directly to the plurality of arithmetic parts formed on one chip through pins fixedly allocated thereto. In contrast, an arrangement of the SIMD parallel computer of this invention is that each arithmetic part is connected to the memory via an address changeover circuit and a data changeover circuit as well. One arithmetic part is selectively connected to the memory under the selection control handled by the central control circuit.

In the thus configured SIMD parallel computer, where a plurality of arithmetic parts are packaged on one chip, the address can be generated in each of the arithmetic units. The address generated by each arithmetic part and the data transferred between the arithmetic part and the memory are inputted and outputted in a time-share manner between the chip and the memory outside the chip through the address changeover circuit and the data changeover circuit. Therefore, the memory access can be carried out with a sufficient data width by using all the pins provided on the chip.

In the parallel computer according to the present invention for performing parallel processing of the data through communications between the plurality of processors arrayed in grids, each processor has output links extending two directions of a first diagonal line of the grid and also input links extending in two directions of a second diagonal line thereof. The output links are connected to input or output links of 3-neighbor processors. Provided is a communication circuit for selecting one of the bidirectional output links by 1 bit of a 2-bit direction signal given to establish the communication direction and one of the bidirectional input links by another 1-bit.

In the computer system of this invention exhibiting such characteristics, when outputting the data to the 3-neighbor processors of the grid, the communication circuit specifies one of the bidirectional output links of the first diagonal line in the right upward (or left upward) direction by 1-bit of the 2-bit direction signal. The communication circuit also specifies one of the bidirectional input links of the second diagonal line in the right downward (or left downward) direction by another 1-bit of the direction signal. In this way, the communications with the grid-4-neighbor processors are attainable simply with the communication circuit specifying two output links and two input links.

BRIEF DESCRIPTION OF THE DRAWINGS

Other objects and advantages of the present invention will become apparent during the following discussion taken in conjunction with the accompanying drawings, in which:

FIG. 1 is a block diagram illustrating a configuration of arithmetic units of an SIMD parallel computer of the present invention;

FIG. 2 is a block diagram showing one example of architecture of the SIMD parallel computer of this invention;

FIG. 3 is a schematic diagram illustrating a data storage format within a memory in a conventional SIMD parallel computer;

FIG. 4 is a schematic diagram illustrating a data storage format within a memory in the SIMD parallel computer of this invention;

FIG. 5 is a block diagram depicting a connection network of the parallel computer of this invention;

FIG. 6 is a block diagram showing a construction of a communication circuit of the processor of the parallel computer of this invention;

FIGS. 7A and 7B are block diagrams showing actual examples of connection networks of the parallel computer of this inventions;

FIG. 8 is a schematic block diagram illustrating an entire architecture of a conventional parallel computer;

FIG. 9 is a block diagram showing a construction of a conventional arithmetic unit;

FIGS. 10 to 12 are block diagrams each showing an example of configuration of an arithmetic unit and an external memory of the conventional SIMD parallel computer;

FIG. 13 is a block diagram illustrating an NEWS type connection network of the conventional parallel computer;

FIG. 14 is a block diagram showing a construction of a communication circuit of the processor of the conventional parallel computer; and

FIG. 15 is a block diagram illustrating an X-Net type connection network of the conventional parallel computer.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Embodiments of the present invention will hereinafter be described with reference to the accompanying drawings.

FIG. 1 is a block diagram illustrating a configuration of an arithmetic unit 102 of an SIMD parallel computer according to the present invention. Note that the whole configuration is the same with the conventional example shown in FIG. 2, and hence the description will be omitted.

An instruction INS given from a control circuit 101 (see FIG. 8) via an instruction bus 112 is imparted to one end of a selector 126 serving as a selecting means. Given to the other end of the selector 126 is an instruction INSP stored in an instruction register 127 conceived as an instruction storage means. The instruction register 127, which is provided in a register group 121 which will be mentioned latter, stores a local instruction INSP generated by an arithmetic element 122. The selector 126 selects one of the two instructions INS and INSP given thereto. The selected instruction INS or INSP is sent to the arithmetic element 122. A selecting operation of the selector 126 is controlled by an instruction selecting circuit 125 defined as a selection control means. Given via a control line 111 to the instruction selecting circuit 125 are a control signal CTR transmitted from the control circuit 101 and an operation authorizing flag which will hereinafter be stated. A changeover signal CS is outputted to the selector on the basis of statuses of the control signal CTR and the operation authorizing flag.

The arithmetic element 122 performs an arithmetic process corresponding to the instruction given. For instance, when storing the instruction INSP in the instruction register 127, a corresponding instruction INS is issued from the control circuit 101 via the selector. Data is read from a local memory 123. An instruction INSP required therefor is generated and stored in the instruction register 127 of the register group 121.

On the other hand, an address ADD imparted from the control circuit 101 via an address bus 113 is given to the local memory 123 or the register group 121. The local memory 23 works to store arithmetic data employed in the arithmetic element 122 and data on the arithmetic results. The register group 121 temporarily stores the arithmetic data stored in the local memory 123 and the data in the middle of arithmetic process and imparts the data to the arithmetic element 122.

The register group 121 has an area for storing operation authorizing flags 124 indicating an authorization of operation, a non-authorization and, if authorized, the alternative of instructions which is selected by the selector 126 depending on the statuses to the arithmetic unit 102. The register group 121 is, as described above, provided with the instruction register 127 for storing the local instruction INSP generated by the arithmetic element 122.

In the thus configured single instruction parallel computer of this invention, the respective arithmetic units 102, 102 . . . execute the common instruction INS issued from the control circuit 101 depending on the statuses of the operation authorizing flags 124 and the control signals CRT transmitted form control circuit 101. Those arithmetic units alternatively execute, after the selector 126 has made a selection whether to execute the local instruction INSP stored in the instruction register 127, the thus selected instruction.

Where addition and subtraction are now simultaneously executed by parallel processing, it is assumed that, for example, an add instruction be a common instruction INS, while a subtract instruction be a local instruction INSP. The subtract instruction is at first generated. For this purpose, the control circuit 101 imparts an instruction to generate the subtract instruction to each of the arithmetic units 102, 102 . . . . At this moment, the operation authorizing flags 124 of the arithmetic units 102, 102 . . . assume an authorized status. In response to the control signal CTR, the instruction selecting circuit 125 outputs the changeover signal CS to cause the selector 126 to select the instruction INS. Outputted at this time from the control circuit 101 via the address bus 113 is the address ADD of such a portion that the data to generate the subtract instruction is stored in the local memory 123. Simultaneously, an address of the instruction register 127 is outputted to the register group 121.

The arithmetic element 122 generates the subtractive instruction which is in turn stored in the instruction register 127. Values distinguishably indicating an authorization, a non-authorization and, if authorized, the selection by the selector 126 are subsequently set to the operation authorizing flag 124 in response to the control signal CTR. When executing the addition and subtraction, all the arithmetic units 102, 102 . . . are brought into the authorized status. Set to the operation authorizing flag 124 are such values as to change over the selector to the control circuit 101 with respect to the arithmetic units 102, 102 . . . which execute the add instruction and to the instruction register 127 with respect to the arithmetic units 102, 102 . . . which execute the subtractive instruction. Next, the control signal CTR is given to the instruction selecting circuit 125. Determined is any one of such three instruction executing patterns that all the arithmetic units 102, 102 . . . execute the local instruction INSP of the instruction register 127; or the arithmetic units 102, 102 . . . execute the instruction INS issued from the control circuit 101; or in accordance with the operation authorizing flag, the arithmetic units execute the instruction INSP of the instruction register 127 on one hand and also execute the instruction INS given from the control circuit 101 on the other hand.

In this case, the third pattern is determined according to the control signal CTR. The values described above are set to the operation authorizing flag 124. The add instruction is issued from the control circuit 101, while the subtract instruction is issued from the instruction register 127.

As explained earlier, according to the present invention, an intrinsic instruction is executable per arithmetic unit in addition to the same single instruction as the conventional one.

The instruction selecting circuit 125 is possible of decision on whether the instruction to be executed is obtained from the common instruction or from the intrinsic instruction register 127. A frequency at which the instructions are generated can be reduced owing to this selection. Namely, when selectively executing the two instructions, one instruction is obtained from the common instruction, and it follows that only the other instruction may be generated for storage in the instruction register 127.

As discussed above, according to the present invention, the arithmetic unit incorporates the instruction storage means for storing the local instruction, the selecting means for selecting any one of the common instruction and the local instruction and the selection control means for controlling the selecting operation of the selecting means. The instruction to be executed by the selecting means is selectable. Hence, there are exhibited excellent advantages in which a plurality of instructions can be simultaneously executed in parallel while holding the characteristics of the single instruction parallel computer, and the processing time can be reduced.

FIG. 2 is a block diagram illustrating one example of configurations of the arithmetic unit and the external memory of an SIMD parallel computer according to the present invention. The same or corresponding components to those of FIGS. 10 through 12 showing the above-described conventional example are marked with the like reference symbols.

Referring to FIG. 2, the reference numeral 202 represents an LSI chip formed with a plurality of arithmetic parts 211, address generating circuits 213 attached to the respective arithmetic parts 211, a single address changeover circuit 221 and a single data changeover circuit 222.

Designated at 212 is a single piece of memory externally attached to one LSI chip 202.

Indicated at 203 is a central control circuit provided for supplying a common memory address and instruction to all the arithmetic parts 211 in the single instruction parallel computer of this invention. The instruction is issued from this central control circuit 203 to the respective arithmetic parts 211, and an address is also given to an address line AL which will be mentioned latter.

The address generating circuits 213 are attached to the respective arithmetic parts 211 and, to be specific, constructed of latch circuits. Each address generating circuit 213 is provided for the purpose of permitting the respective arithmetic parts 211 to refer to the memory 212 at different addresses. More specifically, the individual arithmetic parts 211 normally refer to the memory 212 at the address common to all the arithmetic parts 211 which is outputted from the central control circuit 203. However, each address generating circuit 213 makes each arithmetic part 211 accessible to the memory 212 with an arithmetic result serving as a memory address as the necessity arises.

An address pin and a data pin serve to connect the address changeover circuit 221 and the data changeover circuit 222 on the LSI chip 202 to the memory 212 provided outwardly of the LSI chip 202. These pins are selectively used by one arithmetic part 211 with a changeover to each arithmetic part 211 on the basis of a time-shared usage. A concrete explanation will be given as follows.

The address changeover circuit 221 is concretely composed of a selector and singly provided on each LSI chip 202. The address changeover circuit 221 is connected via the address line AL to the respective arithmetic parts 211 on the LSI chip 202. Note that an address signal line led from the central control circuit 203 is also connected to the address line AL on the LSI chip 202.

The address changeover circuit 221 is connected to the memory 212 through the address pin for connecting the LSI chip 202 to the outside. The address changeover circuit 221 selects one arithmetic part 211 which actually outputs the address to the memory 212 among the plurality of the arithmetic parts 221 on the LSI chip 202. The selecting operation thereof is controlled by a selecting signal outputted from the central control circuit 203.

The data changeover circuit 222 is concretely composed of a selector as in the same way with the address changeover circuit 221. This circuit 222 is singly provided on each LSI chip 202. The data changeover circuit 222 is connected via a data line DL to the individual arithmetic parts 211 on each LSI chip 202.

The data changeover circuit 222 is connected to the memory 212 through the data pin for connecting the LSI chip 202 to the outside. The data changeover circuit 222 selects one arithmetic part 211 which actually outputs the data to the memory 212 and inputs the data from the memory 212 among the plurality of arithmetic parts 211 on the LSI chip 202. The selecting operation thereof is conditioned by address/data selecting signals outputted from the central control circuit 203. As a matter of course, however, the arithmetic part 211 selected by the address changeover circuit 221 is identical with the arithmetic part 211 selected by the data changeover circuit 222.

Based on the architecture described above, sufficient widths can be given to both the address pin and the data pin.

As discussed above, in the great majority of cases a data width of the memory 212 has hitherto been 1 bit in the prior art. In accordance with the architecture of the present invention, however, sufficient bit widths for connection to the memory 212 can be given to the address pin and the data pin as well. Hence, in the prior art, as shown by hatching in the schematic diagram of FIG. 3, the data (4-bit data are shown as one example) which are to be stored in the memory 212 are stored in series astride of a plurality of addresses of the memory 212. By adopting the architecture of this invention, however, as shown by hatching in FIG. 4, the data are stored in parallel in one address. With this arrangement, a correspondence of the storage contents of the memory 212 to the addresses is the same as that of an ordinary appliance. Therefore, a direct reference by the external appliance to the memory 212, i.e., the control over direct memory access (DMA), is facilitated.

Note that in the embodiment discussed above the LSI chip is provided with the address changeover circuit 221 for changing over the address outputted from the central control circuit 203 to the address generated within the LSI chip 202, but this circuit 221 may be, as a matter of course, provided outwardly of the LSI chip 202.

The above-described architecture of the present invention can be used in combination with the conventional architecture previously explained.

As discussed above, according to the present invention, where the plurality of arithmetic parts of the SIMD parallel computer are packaged on one chip, the addresses can be generated in each arithmetic unit. Besides, there are exhibited excellent effects in which no decline in performance can be seen, and a matching property of address with the external appliance is improved.

FIG. 5 is a block diagram illustrating a connection network of the parallel computer of this invention. Individual processors 501, 501 . . . arrayed in grids include and output link 508ne extending in one right upward direction (north east which will hereinafter be abbreviated NE) of two directions of a first diagonal line and an output link 508sw extending in the other left downward direction (south west abbreviated to SW) thereof. Provided also are an input link 508se extending in one right downward direction (south east abbreviated to SE) of two directions of a second diagonal line and an input link 508nw extending in the other left upward direction (north west abbreviated to NW) thereof. The NE-direction output link 508ne of the processors 501, 501 . . . is connected to the SE-direction input link 508se, the SW-direction output link 508sw and the NW-direction input link 508nw of the processors 501, 501 . . . in an upward (hereinafter referred to as N) direction, in the NE-direction and in a right (referred to as E) direction.

It is to be noted that the connections of the respective input/output links are effected based on wired AND. Hence, the signals on the communication links are all negative logic, thereby actualizing wired OR. Other SW-direction output link 508sw, the NW-direction input link 508nw and the SE-direction input link 508se are similarly connected. Therefore, a single link may suffice for one processor 501. A hardware quantity of the communication links is reduced by a factor of 2 as compared with the prior art.

FIG. 6 is a block diagram depicting a configuration of the processors 501, 501 . . . . The processors 501, 501 . . . include arithmetic elements 502 and communication circuits 503. With respect to the data received, the arithmetic element 502 effects a process pursuant to an instruction outputted from an unillustrated control circuit for handling the whole control. The arithmetic element 502 transmits the processed data. The control circuit sends 2-bit direction signals DS to the respective processors 501, 501 . . . . The 2-bit direction signals DS, as shown in Table 2, indicate communication directions of the processors 501, 501 . . . . For instance, when DS=00, it indicates a direction from N to S. Relations between the direction signals and the directions are the same with the conventional example.

              TABLE 2______________________________________Directional signal            Direction______________________________________00               N → S01               E → W10               W → E11               S → N______________________________________

High-order 1 bit of the direction signal is inversioninputted to an output buffer 504a and is at the same time inputted to one end of an output buffer 504b. The output buffers 504a and 504b are constructed by use of open collector type NAND gates. The data transmitted from the arithmetic elements 502 are supplied respectively to the other ends thereof. The outputs of these buffers are given to the NE-direction output link 508ne and the SW-direction output link 508sw, respectively. Low-order 1 bit of the direction signal DS is inputted to one end of an input buffer 505a which employs an AND gate and is simultaneously inversion-inputted to one end of an input buffer 505b. Inversion-inputted to the other ends of the input buffers 505a and 505b are the receive data given from the NW-direction input link 508nw and the SE-direction input link 508se. The outputs of these buffers are supplied via an OR gate 506 to the arithmetic element 502.

One of the output buffers 504a and 504b is selected by the high-order 1 bit of the direction signal DS. One of the input buffers 505a and 505b is selected by the low-order 1 bit. Hence, a decoder required for the prior art is not needed. Besides, the number of gates for the buffers is halved, and the hardware quantity of the communication circuits 503 is also reduced.

The following is a description of the thus constructed parallel computer of this invention. When performing the communications in, e.g., an S→N direction, the direction signal defined by DS=00 is supplied to each of the processors 501, 501 . . . from the control circuit. The output buffers 504a and the input buffers 505b are thereby selected. Only the NE-direction output link 508ne is possible of data output in the respective processors 501, 501 . . . , whereas only the SE-direction input link 508se is impossible of data input. In this status, the NE-direction output link 508ne is connected to the SE-direction input link 508se of the N-direction processor 5-1, as a result of which the communication in the N-direction is established.

The arrangement is the same with other directions such as E→W, W→E and S→N. The communications in four directions are established by a combination of two output links with two input links.

Note that 1-bit width communication links are shown in the architecture of this embodiment. This arrangement is, however, provided for the sake of explanatory convenience. The present invention is not limited to this arrangement but may be, as a matter of course, likewise applicable to multi-bit width communication links.

The connection network in this embodiment involves the use of wired AND. The present invention is not, however, limited to the wired AND. A three-state buffer is usable as an output buffer.

Applicable also to the present architecture are a variety of conventional grid terminal connecting processes such as a torus connection where the processors at bilateral and vertical ends are connected, a cylindrical connection where the processors at bilateral or vertical ends are connected and so on.

As explained earlier, according to the present invention, the grid-4-neighbor (vertical and bilateral) processors are communicable simply by individually selecting two bidirectional output links of the first diagonal line of the grid and two bidirectional input links of the second diagonal line thereof. Exhibited is such an excellent effect that the bidirectional communications with grid-4-neighbor processors are attainable as a function equivalent to the conventional one with a simple construction requiring a less amount of hardware of the communication links and circuits.

FIG. 7A is a block diagram showing an example where the connecting structure of the processors according to the present invention is put into practice on the actual LSI chip. Referring to FIG. 7A, the numeral 701 denotes an LSI chip on which four pieces of processor elements PE0, PE1, PE2 and PE3 are formed. Outputs (NE, SW) of each processor are connected via a driving circuit 702 to a predetermined conductor. Note that this driving circuit 702 may serve as a gate for controlling a signal transfer. In this embodiment, the connections of the four intra-LSI processors are attained by connecting SW of the processor PE1 to NW of the processor PE2 through an OR circuit 703 and further connecting SE of the processor PE0 to NW of the processor PE3.

FIG. 7B is a block diagram illustrating a connecting structure by use of the LSI chip. A large-scale connecting structure according to the present invention can be actualized by connecting, as depicted in the Figure, an E-terminal to a W-terminal, an N-terminal to an S-terminal, an NE-terminal to an SW-terminal and an NW-terminal to an SE-terminal on each LSI chip.

Although the illustrative embodiments have been described in detail with reference to the accompanying drawings, it is to be understood that the present invention is not limited to those embodiments. Various changes or modifications may be effected therein by one skilled in the art without departing from the scope or spirit of the invention.

Claims (7)

What is claimed is:
1. A single instruction parallel computer comprising:
a plurality of arithmetic units;
a control unit for providing the same instruction representing an arithmetic operation, the same address signal, and the same control signals to each of said plurality of arithmetic units to control operation of said arithmetic units, wherein each of said arithmetic units comprises:
(a) an arithmetic element for executing a plurality of arithmetic operations;
(b) a local memory for storing an arithmetic result given by said arithmetic element and data to be computed at said arithmetic element, said local memory being accessed by said address signal from said control unit;
(c) instruction storage means for storing a local instruction representing a local arithmetic operation;
(d) operation authorization storage means for storing an operation authorizing flag representing whether or not said arithmetic element should execute said local instruction; and
(e) instruction selecting means responsive to said control signal from said control unit and said operation authorizing flag read out from said operation authorization storage means, for selecting one of said instruction from said control unit and said local instruction read out from said instruction storage means, and for providing the selected instruction to said arithmetic element to execute the arithmetic operation corresponding to the selected instruction.
2. A single instruction parallel computer as set forth in claim 1, wherein said instruction storage means and operation authorization storage means further comprise:
registers in a register group for temporarily storing the data to be computed by said arithmetic element and the arithmetic result of the computed data.
3. A single instruction parallel computer as set forth in claim 1, wherein said local instruction stored in said instruction storage means is generated by said arithmetic element using instruction generative data stored in said local memory.
4. A single instruction parallel computer as set forth in claim 1, wherein said instruction selecting means further comprises:
selecting means connected to receive said instruction from said control unit and said local instruction read out from said instruction storage means and to provide one of them to said arithmetic element; and
an instruction selection control unit connected to receive said operation authorizing flag read out from said operation authorization storage means and to receive said control signal from said control unit, and to provide a selection control signal to said selecting means.
5. A single instruction parallel computer comprising:
(a) a control unit for providing an instruction, a first address and a selecting signal;
(b) a plurality of arithmetic units provided on a single integrated circuit chip, each of which includes an arithmetic part and an address generating circuit for outputting one of said first address received from said control unit and a second address generated at said address generating circuit;
(c) a memory provided separately from said integrated circuit chip;
(d) address switch means for selecting and providing said second address output from said address generating circuit in one of said plurality of arithmetic units to said memory in a time-shared manner in response to said selecting signal from said control unit; and
(e) data switch means for selectively providing data read out from said memory to said arithmetic part in one of said plurality of arithmetic units, said data switch means operating synchronously with said address switch means.
6. A single instruction parallel computer as set froth in claim 5, wherein said data switch means and said address switch means are provided on said integrated circuit chip.
7. A single instruction parallel computer as set forth in claim 6, further comprising:
at least one additional integrated circuit chip identical to said single integrated circuit chip, and at least one additional memory interconnected with said at least one additional integrated circuit chip.
US08251651 1990-08-08 1994-05-31 Arithmetic unit for SIMD type parallel computer Expired - Fee Related US5418915A (en)

Priority Applications (6)

Application Number Priority Date Filing Date Title
JP21192390A JPH0496189A (en) 1990-08-08 1990-08-08 Single instruction type parallel computer
JP2-211923 1990-08-08
JP26257490A JPH04138582A (en) 1990-09-29 1990-09-29 Single instruction type parallel computer
JP2-262574 1990-09-29
US74156891 true 1991-08-07 1991-08-07
US08251651 US5418915A (en) 1990-08-08 1994-05-31 Arithmetic unit for SIMD type parallel computer

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US08251651 US5418915A (en) 1990-08-08 1994-05-31 Arithmetic unit for SIMD type parallel computer

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
US74156891 Continuation 1991-08-07 1991-08-07

Publications (1)

Publication Number Publication Date
US5418915A true US5418915A (en) 1995-05-23

Family

ID=27329300

Family Applications (1)

Application Number Title Priority Date Filing Date
US08251651 Expired - Fee Related US5418915A (en) 1990-08-08 1994-05-31 Arithmetic unit for SIMD type parallel computer

Country Status (1)

Country Link
US (1) US5418915A (en)

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5630161A (en) * 1995-04-24 1997-05-13 Martin Marietta Corp. Serial-parallel digital signal processor
US5666298A (en) * 1994-12-01 1997-09-09 Intel Corporation Method for performing shift operations on packed data
US5805874A (en) * 1993-03-31 1998-09-08 Motorola Inc. Method and apparatus for performing a vector skip instruction in a data processor
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US5903771A (en) * 1996-01-16 1999-05-11 Alacron, Inc. Scalable multi-processor architecture for SIMD and MIMD operations
US5940592A (en) * 1995-03-03 1999-08-17 Hitachi, Ltd. Communication processing method in parallel computer and system therefor
US6006316A (en) * 1996-12-20 1999-12-21 International Business Machines, Corporation Performing SIMD shift and arithmetic operation in non-SIMD architecture by operation on packed data of sub-operands and carry over-correction
US6052705A (en) * 1995-08-30 2000-04-18 Sony Corporation Video signal processor with triple port memory
US6079008A (en) * 1998-04-03 2000-06-20 Patton Electronics Co. Multiple thread multiple data predictive coded parallel processing system and method
US6266758B1 (en) 1997-10-09 2001-07-24 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US20020116428A1 (en) * 2001-02-21 2002-08-22 Morten Stribaek Polynomial arithmetic operations
US20020116432A1 (en) * 2001-02-21 2002-08-22 Morten Strjbaek Extended precision accumulator
US20040215681A1 (en) * 1994-12-01 2004-10-28 Lin Derrick Chu Method and apparatus for executing packed shift operations
US20040252547A1 (en) * 2003-06-06 2004-12-16 Chengpu Wang Concurrent Processing Memory
US6834337B1 (en) 2000-09-29 2004-12-21 International Business Machines Corporation System and method for enabling multiple signed independent data elements per register
US7039906B1 (en) 2000-09-29 2006-05-02 International Business Machines Corporation Compiler for enabling multiple signed independent data elements per register
US7197625B1 (en) 1997-10-09 2007-03-27 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US20090013151A1 (en) * 2007-07-04 2009-01-08 Kazuhiko Hara Simd type microprocessor
US7599981B2 (en) 2001-02-21 2009-10-06 Mips Technologies, Inc. Binary polynomial multiplier
US7617388B2 (en) 2001-02-21 2009-11-10 Mips Technologies, Inc. Virtual instruction expansion using parameter selector defining logic operation on parameters for template opcode substitution
US20090282223A1 (en) * 2008-05-07 2009-11-12 Lyuh Chun-Gi Data processing circuit
US20090319767A1 (en) * 2008-06-20 2009-12-24 Renesas Technology Corp. Data processing apparatus
US8949575B2 (en) 2011-12-14 2015-02-03 International Business Machines Corporation Reversing processing order in half-pumped SIMD execution units to achieve K cycle issue-to-issue latency
US20150095542A1 (en) * 2013-09-28 2015-04-02 Allan D. Knies Collective communications apparatus and method for parallel systems

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4373183A (en) * 1980-08-20 1983-02-08 Ibm Corporation Bus interface units sharing a common bus using distributed control for allocation of the bus
US4835729A (en) * 1985-12-12 1989-05-30 Alcatel Usa, Corp. Single instruction multiple data (SIMD) cellular array processing apparatus with on-board RAM and address generator apparatus
US4891787A (en) * 1986-12-17 1990-01-02 Massachusetts Institute Of Technology Parallel processing system with processor array having SIMD/MIMD instruction processing
US4896256A (en) * 1986-05-14 1990-01-23 Kabushiki Kaisha Toshiba Linking interface system using plural controllable bidirectional bus ports for intercommunication amoung split-bus intracommunication subsystems
US4992933A (en) * 1986-10-27 1991-02-12 International Business Machines Corporation SIMD array processor with global instruction control and reprogrammable instruction decoders
US5157758A (en) * 1989-11-18 1992-10-20 Thorn Emi Plc Tungsten halogen lamp
US5179714A (en) * 1988-10-07 1993-01-12 Martin Marietta Corporation Parallel bit serial data processor
US5206937A (en) * 1989-08-04 1993-04-27 Matsushita Electric Industrial Co., Ltd. Data communication system having plural stations connected to a single transmission path
US5212777A (en) * 1989-11-17 1993-05-18 Texas Instruments Incorporated Multi-processor reconfigurable in single instruction multiple data (SIMD) and multiple instruction multiple data (MIMD) modes and method of operation
US5218709A (en) * 1989-12-28 1993-06-08 The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration Special purpose parallel computer architecture for real-time control and simulation in robotic applications
US5305462A (en) * 1986-09-18 1994-04-19 Digital Equipment Corporation Mechanism for broadcasting data in a massively parallell array processing system
US5317734A (en) * 1989-08-29 1994-05-31 North American Philips Corporation Method of synchronizing parallel processors employing channels and compiling method minimizing cross-processor data dependencies
US5327538A (en) * 1990-09-11 1994-07-05 Canon Kabushiki Kaisha Method of utilizing common buses in a multiprocessor system
US5347450A (en) * 1989-01-18 1994-09-13 Intel Corporation Message routing in a multiprocessor computer system

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4373183A (en) * 1980-08-20 1983-02-08 Ibm Corporation Bus interface units sharing a common bus using distributed control for allocation of the bus
US4835729A (en) * 1985-12-12 1989-05-30 Alcatel Usa, Corp. Single instruction multiple data (SIMD) cellular array processing apparatus with on-board RAM and address generator apparatus
US4896256A (en) * 1986-05-14 1990-01-23 Kabushiki Kaisha Toshiba Linking interface system using plural controllable bidirectional bus ports for intercommunication amoung split-bus intracommunication subsystems
US5305462A (en) * 1986-09-18 1994-04-19 Digital Equipment Corporation Mechanism for broadcasting data in a massively parallell array processing system
US4992933A (en) * 1986-10-27 1991-02-12 International Business Machines Corporation SIMD array processor with global instruction control and reprogrammable instruction decoders
US4891787A (en) * 1986-12-17 1990-01-02 Massachusetts Institute Of Technology Parallel processing system with processor array having SIMD/MIMD instruction processing
US5179714A (en) * 1988-10-07 1993-01-12 Martin Marietta Corporation Parallel bit serial data processor
US5347450A (en) * 1989-01-18 1994-09-13 Intel Corporation Message routing in a multiprocessor computer system
US5206937A (en) * 1989-08-04 1993-04-27 Matsushita Electric Industrial Co., Ltd. Data communication system having plural stations connected to a single transmission path
US5317734A (en) * 1989-08-29 1994-05-31 North American Philips Corporation Method of synchronizing parallel processors employing channels and compiling method minimizing cross-processor data dependencies
US5212777A (en) * 1989-11-17 1993-05-18 Texas Instruments Incorporated Multi-processor reconfigurable in single instruction multiple data (SIMD) and multiple instruction multiple data (MIMD) modes and method of operation
US5157758A (en) * 1989-11-18 1992-10-20 Thorn Emi Plc Tungsten halogen lamp
US5218709A (en) * 1989-12-28 1993-06-08 The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration Special purpose parallel computer architecture for real-time control and simulation in robotic applications
US5327538A (en) * 1990-09-11 1994-07-05 Canon Kabushiki Kaisha Method of utilizing common buses in a multiprocessor system

Cited By (48)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5805874A (en) * 1993-03-31 1998-09-08 Motorola Inc. Method and apparatus for performing a vector skip instruction in a data processor
US5666298A (en) * 1994-12-01 1997-09-09 Intel Corporation Method for performing shift operations on packed data
US5818739A (en) * 1994-12-01 1998-10-06 Intel Corporation Processor for performing shift operations on packed data
US7480686B2 (en) 1994-12-01 2009-01-20 Intel Corporation Method and apparatus for executing packed shift operations
US20040215681A1 (en) * 1994-12-01 2004-10-28 Lin Derrick Chu Method and apparatus for executing packed shift operations
US5940592A (en) * 1995-03-03 1999-08-17 Hitachi, Ltd. Communication processing method in parallel computer and system therefor
US5630161A (en) * 1995-04-24 1997-05-13 Martin Marietta Corp. Serial-parallel digital signal processor
US6052705A (en) * 1995-08-30 2000-04-18 Sony Corporation Video signal processor with triple port memory
US5903771A (en) * 1996-01-16 1999-05-11 Alacron, Inc. Scalable multi-processor architecture for SIMD and MIMD operations
US6006316A (en) * 1996-12-20 1999-12-21 International Business Machines, Corporation Performing SIMD shift and arithmetic operation in non-SIMD architecture by operation on packed data of sub-operands and carry over-correction
CN1108558C (en) * 1996-12-20 2003-05-14 国际商业机器公司 Single-order multi-data correction circuit facing to arithmetic/shift operation
US8074058B2 (en) 1997-10-09 2011-12-06 Mips Technologies, Inc. Providing extended precision in SIMD vector arithmetic operations
US20020062436A1 (en) * 1997-10-09 2002-05-23 Timothy J. Van Hook Method for providing extended precision in simd vector arithmetic operations
US20110055497A1 (en) * 1997-10-09 2011-03-03 Mips Technologies, Inc. Alignment and Ordering of Vector Elements for Single Instruction Multiple Data Processing
US7793077B2 (en) 1997-10-09 2010-09-07 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US6266758B1 (en) 1997-10-09 2001-07-24 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US20070250683A1 (en) * 1997-10-09 2007-10-25 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US7546443B2 (en) 1997-10-09 2009-06-09 Mips Technologies, Inc. Providing extended precision in SIMD vector arithmetic operations
US20090249039A1 (en) * 1997-10-09 2009-10-01 Mips Technologies, Inc. Providing Extended Precision in SIMD Vector Arithmetic Operations
US7159100B2 (en) 1997-10-09 2007-01-02 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US7197625B1 (en) 1997-10-09 2007-03-27 Mips Technologies, Inc. Alignment and ordering of vector elements for single instruction multiple data processing
US6079008A (en) * 1998-04-03 2000-06-20 Patton Electronics Co. Multiple thread multiple data predictive coded parallel processing system and method
US7039906B1 (en) 2000-09-29 2006-05-02 International Business Machines Corporation Compiler for enabling multiple signed independent data elements per register
US6834337B1 (en) 2000-09-29 2004-12-21 International Business Machines Corporation System and method for enabling multiple signed independent data elements per register
US20020178203A1 (en) * 2001-02-21 2002-11-28 Mips Technologies, Inc., A Delaware Corporation Extended precision accumulator
US8447958B2 (en) 2001-02-21 2013-05-21 Bridge Crossing, Llc Substituting portion of template instruction parameter with selected virtual instruction parameter
US20060190519A1 (en) * 2001-02-21 2006-08-24 Mips Technologies, Inc. Extended precision accumulator
US7711763B2 (en) 2001-02-21 2010-05-04 Mips Technologies, Inc. Microprocessor instructions for performing polynomial arithmetic operations
US20020116428A1 (en) * 2001-02-21 2002-08-22 Morten Stribaek Polynomial arithmetic operations
US7225212B2 (en) 2001-02-21 2007-05-29 Mips Technologies, Inc. Extended precision accumulator
US7599981B2 (en) 2001-02-21 2009-10-06 Mips Technologies, Inc. Binary polynomial multiplier
US7617388B2 (en) 2001-02-21 2009-11-10 Mips Technologies, Inc. Virtual instruction expansion using parameter selector defining logic operation on parameters for template opcode substitution
US20020116432A1 (en) * 2001-02-21 2002-08-22 Morten Strjbaek Extended precision accumulator
US7860911B2 (en) 2001-02-21 2010-12-28 Mips Technologies, Inc. Extended precision accumulator
US7181484B2 (en) 2001-02-21 2007-02-20 Mips Technologies, Inc. Extended-precision accumulation of multiplier output
US20040252547A1 (en) * 2003-06-06 2004-12-16 Chengpu Wang Concurrent Processing Memory
US20090013151A1 (en) * 2007-07-04 2009-01-08 Kazuhiko Hara Simd type microprocessor
JP2009015556A (en) * 2007-07-04 2009-01-22 Ricoh Co Ltd Simd type microprocessor
US8219783B2 (en) * 2007-07-04 2012-07-10 Ricoh Company, Ltd. SIMD type microprocessor having processing elements that have plural determining units
US20090282223A1 (en) * 2008-05-07 2009-11-12 Lyuh Chun-Gi Data processing circuit
US7814296B2 (en) * 2008-05-07 2010-10-12 Electronics And Telecommunications Research Institute Arithmetic units responsive to common control signal to generate signals to selectors for selecting instructions from among respective program memories for SIMD / MIMD processing control
US8402260B2 (en) * 2008-06-20 2013-03-19 Renesas Electronics Corporation Data processing apparatus having address conversion circuit
US20090319767A1 (en) * 2008-06-20 2009-12-24 Renesas Technology Corp. Data processing apparatus
US8949575B2 (en) 2011-12-14 2015-02-03 International Business Machines Corporation Reversing processing order in half-pumped SIMD execution units to achieve K cycle issue-to-issue latency
US8977835B2 (en) 2011-12-14 2015-03-10 International Business Machines Corporation Reversing processing order in half-pumped SIMD execution units to achieve K cycle issue-to-issue latency
US9477628B2 (en) * 2013-09-28 2016-10-25 Intel Corporation Collective communications apparatus and method for parallel systems
US20150095542A1 (en) * 2013-09-28 2015-04-02 Allan D. Knies Collective communications apparatus and method for parallel systems

Similar Documents

Publication Publication Date Title
US3566363A (en) Processor to processor communication in a multiprocessor computer system
US5333280A (en) Parallel pipelined instruction processing system for very long instruction word
US5535408A (en) Processor chip for parallel processing system
US5212773A (en) Wormhole communications arrangement for massively parallel processor
US6345321B1 (en) Multiple-mode memory component
US5440749A (en) High performance, low cost microprocessor architecture
US7321958B2 (en) System and method for sharing memory by heterogeneous processors
US4811413A (en) System of reconfigurable pipelines of generalized neighborhood function morphic image processors
US4739474A (en) Geometric-arithmetic parallel processor
US4992933A (en) SIMD array processor with global instruction control and reprogrammable instruction decoders
US4901224A (en) Parallel digital processor
US4679163A (en) Inverse discrete cosine transform calculation processor
US5742180A (en) Dynamically programmable gate array with multiple contexts
US4704678A (en) Function set for a microcomputer
US5872987A (en) Massively parallel computer including auxiliary vector processor
US6446191B1 (en) Methods and apparatus for efficient synchronous MIMD operations with iVLIW PE-to-PE communication
US6381690B1 (en) Processor for performing subword permutations and combinations
US5551039A (en) Compiling a source code vector instruction by generating a subgrid loop for iteratively processing array elements by plural processing elements
US5247689A (en) Parallel digital processor including lateral transfer buses with interrupt switches to form bus interconnection segments
US4849882A (en) Vector processing system for processing plural vector instruction streams
US6560629B1 (en) Multi-thread processing
US5594918A (en) Parallel computer system providing multi-ported intelligent memory
US5442762A (en) Instructing method and execution system for instructions including plural instruction codes
US4709327A (en) Parallel processor/memory circuit
US5175865A (en) Partitioning the processors of a massively parallel single array processor into sub-arrays selectively controlled by host computers

Legal Events

Date Code Title Description
FPAY Fee payment

Year of fee payment: 4

REMI Maintenance fee reminder mailed
LAPS Lapse for failure to pay maintenance fees
FP Expired due to failure to pay maintenance fee

Effective date: 20030523